Market baskets Frequent itemsets FP growth. Data mining. Frequent itemset Association&decision rule mining. University of Szeged.

Transcription

1 Frequent itemset Association&decision rule mining University of Szeged

2 What frequent itemsets could be used for? Features/observations frequently co-occurring in some database can gain us useful insights How a marketing person can make use of it? What about a data scientist? Transaction ID 1 5 Items {milk, bread, salami} {beer, diapers} {beer, wurst} {beer, baby food, diapers} {diapers, coke, bread}

3 Possible forms Horizontal Transaction ID 1 5 Items {milk, bread, salami} {beer, diapers} {beer, wurst} {beer, baby food, diapers} {diapers, coke, bread}

4 Possible forms Vertical/inverted Item milk bread salami beer diapers wurst baby food coke Basket {1} {1, 5} {1} {,,} {,,5} {} {} {5}

5 Possible forms Relational Basket Item milk bread salami beer diapers beer wurst beer baby food diapers diapers coke bread

6 Collecting association rules Goal: find item sets of the transactional database with high support and confidence Support of set X : s(x ) = {ti ti T X ti } Normalized support: s(x ) normalized by the number of transactions Confidence of rule A B: c(a B) = s(a B) s(a) P(B A) The meaning of a rule A B: given that items included in set A are put in the basket, chances are high that the items included in B are also in some basket

7 Interestingness of association rules Are all rules with high support and confidence equally interesting? Confidence of a rule can simply be high due to the fact the items on its right side are frequently purchased independently of the items on the left side. Any example? Interest score of rule A B: I (A B) = c(a B) P(B) What is the interpretation of this score? (Interesting rules will have high absolute values. Why?) Choose threshold so that the number of rules defined is manageable There are plenty other scores for measuring interestingness: χ, κ,...

8 What frequent itemsets can be utilized for? Plagiarism detection Somewhat counter-intuitively let the sentences be the baskets and documents the items Conclusion Sometimes we need to be flexible about the concept of items comprising baskets Our goal is to examine the relation of items to each other, and not that of baskets (we had that one before) How frequent itemsets could be interpreted if items and baskets were determined vice versa?

9 What else frequent itemsets could be used for? Data mining We can use them to build a simple classifier Missing feature values can be estimated knowing how features typically co-occur with each other We can merge features if they show highly similar behavior in the database

10 The general schema of producing associational rules Collect the set of frequent items F having a (normalized) support surpassing some threshold t Partition F into non-empty, disjunct subsets and calculate the confidence of the rules determined

11 Why naive approach fails? How a naive approach would look like? d items d d possible rules (e.g. d = possibilities) d P d d j Proof (hint: (1 + x)d = + xd) j x j=1 d P i=1 d d i d P i d i j j=1 d = d P i=1 d i (d i 1) = d P i=1 d i d i d P i=1 ( ) (d 1) 106 log10 (Possible association rules) Possible association rules d d d i =

12 Itemset I is frequent J I itemsets are frequent What can we say if itemset I is not frequent? Anti-monotone property: function f is said to be anti-monotone if X, Y P (U) : X Y f (X ) f (Y ) {a} {b} {c} {a,b} {a,c} {b,c} {a,b,c}

13 The in action Let assume a frequency threshold of Item beer bread coke diapers milk wurst Frequency 1

14 The in action Let the frequency threshold be Item Frequency beer bread coke diapers milk wurst 1 Item Frequency {beer, bread} {beer, diapers} {beer, milk} {bread, diapers} {bread, milk} {diapers, milk}

15 Algorithm 1 Pseudocode for calculation of frequent itemsets Input: set of possible items U, transaction database T, frequency threshold t Output: frequent itemsets 1: C1 := U : Calculate the support of C1 : F1 := {x x C1 s(x) t} : for (k = ; k < U &&Fk 1 6= ; k + +) do 5: Determine Ck based on Fk 1 6: Calculate the support of Ck 7: Fk := {X X Ck s(x ) t} 8: end for 9: return ki=1 Fi

16 Possible ways of determining Ck Based on the elements in F1. Avoid this! Combining the elements of Fk 1 and F1. Somewhat better Combining the elements of Fk 1 and Fk 1 itself. Generates less candidates but does not miss any candidate which has the chance to be a frequent itemset. Combine two itemsets if A = {a1, a,..., ak 1 }, B = {b1, b,..., bk 1 } Fk 1 : ai = bi, 1 i k ak 1 6= bk 1 We need an ordering over the elements of F (e.g. by converting them to integers) and store them in that form

17 Possible ways of calculating the frequency for Ck Fk 1 frequencies might need to be stored memory limitations Depending on the ratio of non-zero elements, we can either store them explicitly or store them in the form of (i, j, c) triplets, meaning that item i and j has a co-occurrence frequency of c In a one-dimensional triangular matrix the frequency of item pair (i, j), i j is stored at index (i 1)( Fk 1 i ) + j i If the number of zero frequencies surpasses representation using triplets pays off (i, j, c) ) ( Fk 1

18 Calculating Ck example Suppose there are 107 baskets, 10 items/baskets There are 105 different items in total Using the triangular matrix representation integer is required In worst case there are at most different pairs of items in the transaction dataset max = integer suffices to store the non.zero frequencies of the purchase of item pairs

19 Compressing frequent itemsets Maximal frequent itemsets: {I I frequent J I } Closed itemsets: I : s(j) = s(i )} Closed frequent itemsets: closed itemset having a support above some frequency threshold Can a maximal frequent itemset be non-closed? Can a closed itemset be non-maximal?

20 Compressing frequent itemsets Example (t = ) Item A B C AB AC BC ABC Frequency 5 Maximal Closed Closed frequent

21 Compressing frequent itemsets Example (t = ) Item A B C AB AC BC ABC Frequency 5 Maximal Closed Closed frequent

22 The relation of different classes to each other Closed frequent itemsets Maximal frequent itemsets

23 PCY (Park-Chen-Yu) algorithm an extension to A priori Besides counting the frequency of standalone items keep track of the frequency of buckets into which pairs of elements get assigned according to some hash function What can be said based on the aggregated frequency counts? What cannot be made for sure? Only pairs that consist of frequent items and which got hashed to a frequent bucket have the chance to be indeed frequent in the end

24 Practical considerations We might as well do sampling from the transaction database correctness and completeness is sacrificed Different strategies can be applied for extracting frequent itemsets Frequent itemset border null null {a1,a,...,an} (a) General-to-specific/ Top-down Frequent itemset border null.... {a1,a,...,an} Frequent itemset border (b) Specific-to-general/ Bottom-up {a1,a,...,an} (c) Bidirectional/ Hybrid

25 Generalizing A priori It is not necessary to do the expansion according to the different levels of the item lattice We can do expansion according to any equivalent classes Originally we defined equivalent classes based on the size of the itemsets Alternatively we can define equivalent classes based on the size of the overlap in the prefixes/suffixes

26 Generalizing A priori the prefix and suffix view null A AB ABC B AC AD ABD ACD null C BC BD BCD D CD A AB AC ABC B C BC AD ABD D BD CD ACD BCD ABCD ABCD (a) Prefix tree (b) Suffix tree

27 FP trees Use a data structure which makes the extraction of frequent datasets easy FP trees: alternative, condensed representation It can be constructed by processing the rows of the transaction database in one pass are represented as paths in the tree Certain item subsets show up multiple times overlapping paths compressability Hopefully the whole transaction dataset can be stored in the main memory as a result

28 Constructing FP trees Useful heuristic: order items according to their decreasing support Processing one line of the transaction database Continue paths starting from the root with the same prefix to the currently processed line Increment the frequencies stored for visited nodes Set the frequency of the new nodes to 1 Otherwise start a new path from the root node Assign 1 for the frequency value of new nodes Include pointers between items of the same kind

29 FP tree example 1 1 based on the slides of Tan, Steinbach, Kumar: Introduction to Data Mining

30 FP tree after processing the last basket TID Items {A,B} {B,C,D} {A,C,D,E} {A,D,E} {A,B,C} {A,B,C,D} {B,C} {A,B,C} {A,B,D} {B,C,E} A:7 B: C: B:5 C:1 D:1 D:1 E:1 D:1 C: D:1 D:1 E:1 E:1 What trees would we get if items within baskets were ordered according to their increasing/decreasing order of overall support? ({A:7,B:8,C:7,D:5,E:})

31 FP-Growth algorithm Divide and conquer algorithm working on the FP tree in a bottom-up manner If the FP tree reveals that an itemset is frequent check out its supersets on their frequency Examine FP trees conditioned on some already known frequent itemsets E.g. it turns out {E} is frequent check out the frequency of set {A,E}, {B,E}, {C,E} and {D,E}

32 FP trees conditioned on some target item(s) The part of the FP trees that we would get if we looked at only transactions containing the target item(s) Without building the tree from scratch 1 TID Forget about the parts of the tree not related to the target item(s) Basket {A,B} {B,C,D} {A,C,D,E} {A,D,E} {A,B,C} {A,B,C,D} {B,C} {A,B,C} {A,B,D} {B,C,E} A:7 B: C:1 D:1 C: D:1 E:1 E:1 E:1 FP tree conditioned on {E}

33 FP trees conditioned on some target item(s) The part of the FP trees that we would get if we looked at only transactions containing the target item(s) Without building the tree from scratch TID Let the support of a node be the sum of the upgraded supports of its descendants Basket... {A,C,D,E} {A,D,E}... {B,C,E} A: C:1 D:1 B:1 C:1 D:1 FP tree conditioned on {E}

34 FP trees conditioned on some target item(s) The part of the FP trees that we would get if we looked at only transactions containing the target item(s) Without building the tree from scratch TID Eliminate items below frequency threshold (e.g. using a frequency threshold of ) Create FP trees conditioned on item pairs (containing item {E}), based on which we can determine frequent item triples (e.g. {D,E} {A,D,E}) Basket... {A,C,D,E} {A,D,E}... {B,C,E} A: C:1 C:1 D:1 D:1 FP tree conditioned on {E}